acceptodds
Under review as a conference paper at ICLR 2027

TCPL: Target Consensus-Guided Prototype Learning for Open-Scene Target Detection

Abstract

Spectral small target detection in open scenes is challenged by variations in target spectra and backgrounds, which limit the transferability of fixed target priors and target knowledge derived from labeled scenes and varying references. To address this challenge, we propose Target Consensus-Guided Prototype Learning (TCPL), a framework that learns transferable target knowledge for detection without adaptation to new scenes. During densely supervised training, TCPL uses sparse target points and their neighborhoods to construct local spectral–spatial references. Its asymmetric text guidance combines target and background descriptions on the support side while retaining only target semantics on the query side. By matching features from support-point neighborhoods with query-pixel features, TCPL aggregates predictions for the same query image under different support conditions into a consensus, mitigating the effects of variations among support samples. Detection supervision and knowledge distillation transfer this consensus knowledge into a learnable prototype bank and target anchor, enabling query-only inference. Furthermore, we construct three datasets with pixel-level target annotations and establish a unified open-scene evaluation setting.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Small target detection, Cross-scene generalization, Few-shot learning, Text-guided learning, Prototype learning, Knowledge distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JaHHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TCPL for pixel-level spectral small target detection of known categories in unseen scenes. Training combines sparse support-point references, target/background descriptions, and dense query supervision. ATKE provides asymmetric text guidance, TCSV aggregates predictions under different support conditions into a consensus teacher, and CPLD distills these predictions into a learnable prototype bank and target anchor. At inference, only the query image is required, without support inputs, text encoding, or parameter updates. Experiments on Aircraft-14, Vehicle-74, and Plane-400 report improvements over adapted few-shot segmentation and hyperspectral target detection baselines, together with component and loss ablations.

Critical strengths

(1) Practical query-only deployment. Distilling support-conditioned predictions into learned prototypes and an anchor removes the need for support samples, text inputs, or test-time adaptation. This is a useful deployment property for detecting a known target category in new imagery. Asymmetric text guidance is also reasonably motivated by the potential mismatch between support and query backgrounds.

(2) Meaningful empirical gains and evaluation effort. The paper evaluates three datasets, multiple point budgets, and three random seeds, and includes internal baselines and ablations. On Vehicle-74, full TCPL improves AP from 51.79 to 62.52 relative to the Query-only variant (Table 4), providing evidence beyond comparisons with external methods. The controlled point-budget experiment and additional San Diego testing are useful supplements. The construction of the evaluation datasets, including new pixel-level aircraft annotations for Plane-400, is another potential contribution.

Critical weaknesses

Weakness 1:The baseline selection does not fully establish the advantage in the stated supervised setting. TCPL uses dense training-query masks and detects known categories at test time. The main comparisons contain six adapted few-shot segmentation methods and two hyperspectral target detection methods, but no independent, well-tuned standard supervised segmentation baseline matched to this setting. Appendix B.2 confirms that most FSS baselines also use dense training-query supervision, so this is not simply a dense-versus-sparse supervision comparison; however, MPA uses a different supervision protocol, and the effects of adapting the baselines to spectral inputs and point maps remain insufficiently explained. The strong internal Query-only result on Aircraft-14 (36.75 AP and 97.40 AUC, Table 4) makes an independent supervised baseline particularly important. Without it, the source of the large margins over the adapted methods remains uncertain. Weakness 2: The specific contribution and generalization benefit of consensus are not sufficiently isolated. Removing TCSV changes both attention retrieval to mean pooling and the multi-support teacher to a single teacher (Table 10). This prevents attributing the resulting difference specifically to consensus. Table 5 provides a cleaner KD ablation: KD improves Vehicle-74 AP by 4.94 percentage points, but Aircraft-14 AP by only 0.50, with increased AP standard deviations on both datasets. These results support dataset-dependent mean improvements, rather than a general claim of improved training stability. Evidence under independent scene shifts is also limited: Aircraft-14 has three test images, Vehicle-74 consists of crops from one flight-line image, and the additional San Diego evaluation lacks baseline comparisons. These limitations do not establish leakage or invalidate the reported gains, but leave the central transfer claim incompletely supported.

Questions for authors
  1. Could you provide a standard supervised segmentation or spectral detection baseline trained with the same dense masks, spectral inputs, splits, and model-selection budget, including comparison on the additional San Diego scenes? Please also clarify how unannotated support pixels are treated in the adapted FSS baselines.
  2. With attention retrieval, support exposure, supervised losses, and training budget held fixed, how much does a multi-support teacher improve over a single-support teacher? Please distinguish this effect from the benefit of retrieval and from the KD-versus-no-KD comparison.
  3. How are the Vehicle-74 spatial regions partitioned, and can the source-scene/crop mappings be provided? For Plane-400, does the source-image split also separate geographic locations? These details would clarify the scope of the generalization results.
  4. Could you provide representative generated target/background descriptions and compare instance-specific text with a fixed category description? This would clarify whether the text contributes sample-specific information beyond a category-level prior.
4 comments

Thank you for the detailed feedback.

To Q1:

1. Standard supervised baseline. We added DeepLabV3+ (ResNet-50), trained with the same dense training-query masks, full spectral inputs, splits, and number of epochs as TCPL. Both select checkpoints by validation AP. DeepLabV3+ uses only the query image and requires no support points or text.

TCPL achieves higher mean AP and AUC in all nine dataset–budget combinations. Its AP gains range from 17.00 to 21.86 percentage points on Vehicle-74 and from 5.52 to 6.08 on Plane-400. On Aircraft-14, TCPL leads by 12.69, 0.60, and 5.06 points at 1, 3, and 5 points, respectively.

Table 1. DeepLabV3+ and TCPL (%, mean ± SD)

Dataset / method1-point AP1-point AUC3-point AP3-point AUC5-point AP5-point AUC
Aircraft-14 / DeepLabV3+30.27±2.5988.20±7.0830.27±2.5988.20±7.0830.27±2.5988.20±7.08
Aircraft-14 / TCPL42.96±4.8097.69±0.4130.87±14.6693.54±3.7935.33±3.0796.10±0.46
Vehicle-74 / DeepLabV3+40.66±2.3882.91±2.5640.66±2.3882.91±2.5640.66±2.3882.91±2.56
Vehicle-74 / TCPL62.52±1.5093.73±0.9559.63±3.6793.67±1.5957.66±1.8094.79±2.07
Plane-400 / DeepLabV3+30.51±0.4885.70±4.2430.51±0.4885.70±4.2430.51±0.4885.70±4.24
Plane-400 / TCPL36.59±2.2796.79±0.3036.33±3.5096.60±0.8536.03±3.3396.64±0.85

2. Unseen San Diego scenes. We evaluated the Aircraft-14-trained DeepLabV3+ without fine-tuning. TCPL achieved higher AP and AUC on both scenes.

Table 2. San Diego evaluation (%)

MethodSD1 APSD1 AUCSD2 APSD2 AUC
DeepLabV3+18.0895.5223.2694.13
TCPL36.5097.7024.4096.80

3. Unannotated support pixels. Support-point maps encode annotated foreground as 1 and all unannotated pixels as 0, including unmarked targets. HSNet masks zero-valued support features. DCAMA uses the map as attention values, with zero-valued positions still participating in attention normalization. IFA, DCDNet, and FPTrans build background prototypes or proxies from non-point positions. During downsampling, HSNet, DCAMA, and IFA use adaptive max pooling. DCDNet and FPTrans use it only when it preserves more nonzero cells than their original interpolation. All five compute the segmentation loss using dense training-query masks. MPA supervises augmented support views with correspondingly transformed point maps, counts non-point pixels as background in its loss, and uses no dense masks from independent training queries. Therefore, we report MPA as a reference under its own supervision protocol.

To Q2:

With attention retrieval fixed, the multi-support teacher improved mean AP by 4.71 points on Aircraft-14 and 5.71 on Vehicle-74, and mean AUC by 0.24 and 0.59 points, respectively. Both settings use the same four supports per query, and all four predictions receive detection supervision. The single-support teacher uses the prediction from the support least often selected as teacher for that query, breaking ties by support ID. Both use KD with identical supervised losses and optimization budgets.

Table 3. Retrieval and teacher aggregation (%, mean ± SD)

Retrieval + teacherAircraft-14 APAircraft-14 AUCVehicle-74 APVehicle-74 AUC
Mean pooling + single42.01±1.6197.76±0.3155.40±3.2792.24±2.22
Mean pooling + mean38.53±5.0997.47±0.2655.96±1.9892.19±0.78
Attention + single38.25±4.5897.45±0.1056.81±2.7093.14±1.76
Attention + mean42.96±4.8097.69±0.4162.52±1.5093.73±0.95

With the multi-support teacher fixed, switching from mean pooling to attention retrieval increased mean AP by 4.43 and 6.56 points on Aircraft-14 and Vehicle-74, respectively. The separate KD ablation measures the effect of adding distillation, with mean AP gains of 0.50 and 4.94 points, respectively.

AuthorsOpenReview ↗

Q3–Q4, W1–W2:

To Q3:

  • Vehicle-74 comes from a single AeroRIT flight-line image. We partitioned the image before cropping. In original-image pixels, the training, validation, and test regions span [7,1735)[7,1735), [1735,2247)[1735,2247), and [2247,3975)[2247,3975) along the horizontal axis, with 32, 10, and 32 retained patches, respectively. Within each region, we cropped non-overlapping patches starting at y=53y=53 with a stride of 256 pixels and padded edge patches to 256 × 256 after cropping. The 512-pixel-wide validation region separates the training and test patches, so they are never directly adjacent. We also checked for vehicles crossing subset boundaries and excluded the affected patches. No source pixels or vehicle instances are shared across subsets. Vehicle-74 therefore evaluates spatial generalization within one acquisition.

  • Plane-400 is split by full source-image ID into 20/4/7 source images, corresponding to 264/39/97 patches. The test source images are held out in their entirety, although airport locations overlap with training. To supplement the cross-location evidence, we completed pixel-level annotations during the review period for 167 additional patches from 13 WorldView-3 images covering 10 unseen airports. Neither its airport IDs nor its satellite strip IDs overlap with the original data. These patches were used exclusively for testing.

To Q4:

The fixed category sentences are “An aircraft viewed from above.” and “A road vehicle viewed from above.” Target descriptions add visible instance structure when available. Background descriptions capture the land cover and layout of the same support image.

Table 4. Examples of descriptions used in training

Training supportTarget text beyond the category sentenceBackground text
ABU-Airport-2cross-like silhouette, paired lateral wings, and central body trace.Paved surface bordered by linear roads, a rectangular structure, and mottled patches, possibly vegetation.
AVIRIS Scene 01, group 3Category sentence onlyOpen ground with scattered irregular patches, possibly vegetation, crossed by a narrow diagonal linear road.
Vehicle-74, 03Structural cues: compact rectangular footprint and roof-body trace.large striped roof beside intersecting roads and grass strips with shrubs

We held each support’s background description and all training settings fixed and varied only the target text. Adding instance structure increased mean AP by 1.71 points on Aircraft-14 and 7.29 on Vehicle-74. Mean AUC increased by 0.89 points on Vehicle-74 and was unchanged on Aircraft-14.

Table 5. Target-text ablation (%, mean ± SD)

Target textAircraft-14 APAircraft-14 AUCVehicle-74 APVehicle-74 AUC
Category sentence only41.25±2.2797.69±0.3555.23±3.4292.84±1.44
Category + instance structure42.96±4.8097.69±0.4162.52±1.5093.73±0.95

To W1:

We added an independent DeepLabV3+ baseline under matched dense supervision and clarified how each adapted FSS method uses unannotated support pixels.

To W2:

We added a teacher-aggregation comparison with retrieval and KD fixed, revised the conclusions to distinguish mean performance gains from training stability, and added DeepLabV3+ comparisons on both San Diego scenes.

If any questions remain, please let us know. We will do our best to address them promptly and improve the manuscript.

Dear Reviewer, with less than three days remaining in the discussion period, we would appreciate your review of our rebuttal and your feedback on whether it addresses your concerns. Thanks again for your time and help in improving the quality of this paper.

Dear Reviewer, with less than one day remaining in the discussion period, we would appreciate your review of our rebuttal and your feedback on whether it addresses your concerns. Thanks again for your time and help in improving the quality of this paper.

Reviewer_VHooOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies open-scene spectral small target detection, where a detector trained on labeled scenes must work on unseen scenes without test-scene adaptation. Target Consensus-Guided Prototype Learning (TCPL) is proposed to learn transferable target knowledge through support-conditioned training with dense query supervision for detection in unseen scenes without query-scene adaptation. The authors also introduce Aircraft-14, Vehicle-74, and Plane-400 with pixel-level masks and compare against few-shot segmentation and hyperspectral target-detection baselines.

Critical strengths

The training/inference distinction is well motivated and technically coherent. Support-conditioned matching is used to expose the model to reference variation during training, while consensus distillation transfers that information into prototypes and an anchor, so inference does not require support images, text encoding, or test-time adaptation. This is a useful formulation for deployment in genuinely unseen scenes.

Critical weaknesses
  1. The Introduction lacks a clear explanation of the motivation behind the proposed method. In particular, the connection between the specific challenges being addressed and the design of the corresponding modules is not sufficiently established.

  2. The evidence for open-scene generalization remains limited by the effective number and diversity of independent scenes. Aircraft-14 has only 14 images (9/2/3 split) and Vehicle-74 is cropped from one AeroRIT flight-line image into non-overlapping patches (32/10/32); the latter therefore does not establish generalization across acquisition scenes. Plane-400 is larger, but comes from one satellite product/domain and has only eight bands.

  3. The central claim that support consensus yields transferable knowledge is only partially isolated. The ablations show improvements from the full system, but do not quantify the variability of individual support-conditioned predictions versus the consensus and prototype student across many independent scenes. The headline gains are also uneven: on Aircraft-14 at 3 points the AP difference from SelfMTL is 0.07 pp with a 14.66 standard deviation, and AUC is lower.

Questions for authors
  1. Could the authors further elaborate on the motivation for incorporating external information in the Introduction and explain the underlying mechanisms through which the proposed modules leverage such information to enhance the model robustness across different scenarios?

  2. For each dataset, what are the original acquisition scenes in train/validation/test? In particular, how can Vehicle-74 support an open-scene claim if its patches are all cropped from a single flight-line image covering a campus area?

  3. Can you report per-scene AP and the variance of predictions across support conditions, consensus teacher, and query-only Bank, using identical queries? This would clarify whether consensus distillation specifically reduces support sensitivity rather than the gain coming mainly from extra dense supervision or the prototype architecture.

  4. The prompts are generated from full support images and target crops, with descriptions then inspected manually. Do results hold when descriptions are generated automatically without manual curation?

5 comments
AuthorsOpenReview ↗

Q1–Q2: Method Motivation, Dataset Splits, and Transfer to Unseen Airports

Thank you for the thoughtful and detailed reviews.

To Q1:

We revised the Introduction to explain why local spectral measurements can be ambiguous across scenes and how this motivates TCPL. A target’s spectrum can vary with acquisition conditions, and background materials can have similar spectral responses. To complement the spectrum measured at a support point, we use a pretrained vision-language model to describe the selected target’s visible structure and the layout of its training-support image.

ATKE contrasts target and background descriptions on the support image. For training queries, it uses only the target description because the support and query backgrounds may differ. TCSV matches support-point neighborhoods to query pixels and averages predictions across supports to reduce sensitivity to the choice of reference. CPLD distills this consensus into a prototype bank and target anchor under dense query supervision. At test time, the detector requires only the query image.

To Q2:

  • Aircraft-14. Training uses eight patches from the Mojave AVIRIS scene and ABU-Airport-2 (Los Angeles). Validation uses ABU-Airport-1 (Los Angeles) and ABU-Airport-4 (Gulfport). Testing uses two patches from the Moffett AVIRIS scene and ABU-Airport-3 (Los Angeles). Moffett is an unseen acquisition. ABU-Airport-3 is a held-out image from the Los Angeles ABU data, represented in training by a different image, ABU-Airport-2.
  • Vehicle-74. All patches come from one AeroRIT campus flight-line image. We split the image spatially before cropping, retaining 32/10/32 training/validation/test patches. The 512-pixel-wide validation region separates training from testing. Vehicle-74 therefore evaluates spatial generalization within one acquisition.
  • Plane-400. All crops from a source image belong to the same subset. Training, validation, and testing use 20/4/7 complete WorldView-3 source images, yielding 264/39/97 patches. The seven test images cover four airports also represented by other training images. This split evaluates transfer to unseen imagery at previously represented airports.

To evaluate transfer to unseen airports, we completed pixel-level annotations during the review period for 167 additional patches from 13 WorldView-3 images at 10 new airports. We used these patches exclusively for testing. Their airports and satellite strips are disjoint from the original data. We retained the same eight-band inputs and pixel-label processing. TCPL used frozen Bank weights selected on the original validation set, with no fine-tuning.

Table 1. Frozen evaluation on 10 unseen airports by support-point budget (%, mean ± SD)

Method1-point AP1-point AUC3-point AP3-point AUC5-point AP5-point AUC
HSNet18.12±0.5785.61±1.0521.03±1.0491.72±0.9321.25±0.3992.15±0.17
DCAMA25.92±0.7093.10±0.1329.25±0.4093.17±0.5728.78±0.8093.23±0.18
IFA18.80±0.2890.83±0.2719.14±0.2691.15±0.2218.85±0.0591.44±0.02
DCDNet16.68±0.4786.71±0.4216.79±0.7987.84±2.2218.08±0.8891.45±0.65
FPTrans16.03±0.0887.58±0.1014.60±1.0186.27±0.4112.91±1.5786.86±0.29
MPA10.45±0.4185.64±5.0214.37±2.2987.78±4.7715.24±1.3089.15±3.72
SelfMTL18.44±2.2287.61±4.5319.03±2.4487.56±0.8418.27±2.6487.40±0.72
TCFSL27.94±6.1285.80±3.2223.74±5.7961.41±4.4724.34±4.4458.86±4.45
DeepLabV3+22.34±1.4279.56±4.6822.34±1.4279.56±4.6822.34±1.4279.56±4.68
TCPL28.25±1.6591.98±2.4329.07±3.0293.90±0.3528.33±3.1493.44±1.20

On these unseen airports, frozen TCPL ranked among the top two methods in both mean AP and mean AUC at every budget. It achieved the highest mean AP at 1 point, exceeding TCFSL by 0.31 percentage points, and the highest mean AUC at 3 and 5 points, exceeding DCAMA by 0.73 and 0.21 points, respectively.

AuthorsOpenReview ↗

Q3: Per-Scene AP and Sensitivity to Support Selection

To Q3:

Across five test subsets, the four-support consensus teacher reduced prediction SD by 53.9%–68.4% relative to Online. For each seed, we used the same checkpoint, queries, and masks. Online predicts from one support at a time, the teacher averages probability maps from four supports in the same pool, and Bank uses only the query.

Table 2. Paired 1-point AP (%, mean ± SD over three seeds)

Test subsetOnlineTeacherBank
Aircraft-14 / Moffett AVIRIS40.05±6.2440.49±6.4141.54±5.87
Aircraft-14 / ABU-Airport-345.37±3.3345.70±3.2945.79±3.56
Vehicle-74 / held-out region62.42±1.5462.45±1.5462.52±1.50
Plane-400 / original test set36.50±2.4736.58±2.4136.59±2.28
External / unseen airports28.21±1.5228.29±1.5928.25±1.66

Table 3. Prediction SD across support conditions (percentage points)

Test subsetOnlineTeacherBank
Aircraft-14 / Moffett AVIRIS1.6554±0.66160.5235±0.20920±0
Aircraft-14 / ABU-Airport-31.4308±0.30650.4525±0.09690±0
Vehicle-74 / held-out region0.4614±0.09370.2127±0.05660±0
Plane-400 / original test set0.4160±0.24720.1890±0.11380±0
External / unseen airports0.4767±0.28270.2172±0.13470±0

Tables 2–3 average over query patches. For Table 3, we compute each valid pixel’s prediction SD across support conditions and take the root mean square over pixels. In both tables, ± denotes the sample SD across three training seeds. Bank’s zero support SD follows from query-only inference.

Table 4. Per-source-image AP at 1 point (%, mean over three seeds). Values are Online / Teacher / Bank.

Source image within each setPlane-400 (97 patches, 7 images)External (167 patches, 13 images)
124.72 / 24.77 / 24.7219.86 / 19.92 / 19.87
224.37 / 24.42 / 24.3745.73 / 45.83 / 45.65
35.46 / 5.48 / 5.4644.17 / 44.25 / 44.24
422.49 / 22.60 / 22.4739.97 / 40.07 / 40.01
567.86 / 67.95 / 68.0110.41 / 10.44 / 10.41
655.68 / 55.79 / 55.8916.08 / 16.14 / 16.03
745.33 / 45.41 / 45.5427.82 / 27.91 / 27.82
8—3.15 / 3.34 / 3.46
9—20.82 / 20.90 / 20.81
10—6.91 / 6.94 / 6.99
11—44.46 / 44.55 / 44.59
12—53.19 / 53.27 / 53.18
13—1.67 / 1.67 / 1.65

Table 4 pools valid pixels across all patches from each source image. Applying the same calculation to Moffett’s two patches gives 32.62 / 32.99 / 33.64 AP for Online / Teacher / Bank.

Of the 20 source images in Table 4, the teacher achieved a higher mean AP than Online on 19 and tied on one at the reported precision. Bank scored slightly below the teacher on 14 images and above it on six. Bank stayed within 0.18 AP percentage points of the teacher on every image, using only the query image at inference.

To isolate the training contribution of consensus, we held dense supervision, prototype architecture, attention retrieval, support exposure, and optimization budget fixed. Replacing the single-support teacher with the four-support mean teacher increased mean AP by 4.71 percentage points on Aircraft-14 and 5.71 on Vehicle-74.

AuthorsOpenReview ↗

Q4, W1–W3: Automatic Text Generation and Responses to Weaknesses

To Q4:

The reported experiments already use automatically generated descriptions. We designed and refined prompts shared by all samples. The final descriptions were generated automatically and assembled into training texts using fixed rules. Manual inspection consisted only of reading the outputs. No individual description was edited, filtered, or regenerated. Omitting this check would leave the training texts unchanged.

To W1:

We revised the Introduction to connect spectral ambiguity to asymmetric text guidance, variation among supports to consensus aggregation, and query-only deployment to prototype distillation.

To W2:

We clarified the evaluation scope of each dataset and added a separate frozen-model evaluation on 10 previously unseen airports, including comparisons with the baselines in Table 1.

To W3:

We added the identical-query prediction analysis and controlled teacher comparison in Q3. We also revised the discussion of Aircraft-14 at 3 points. TCPL and SelfMTL have similar mean AP (30.87±14.66% and 30.80±1.53%, respectively), and SelfMTL has higher mean AUC (94.72% versus 93.54%). We now distinguish reduced sensitivity to support choice from variability across training seeds.

If any questions remain, please let us know. We will do our best to address them promptly and improve the manuscript. If the new experiments and clarifications address your concerns, we would appreciate your consideration of a higher rating.

Reviewer_x2oEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies pixel-level spectral target detection for known categories in unseen scenes. TCPL combines asymmetric text guidance, retrieval from sparse support-point neighborhoods, and a teacher formed by averaging predictions across multiple support conditions. Dense query supervision and output distillation train a prototype bank and target anchor, enabling query-only inference without support inputs, text encoding, or test-time parameter updates. Experiments on three constructed datasets report comparisons across support-point budgets, component ablations, and additional qualitative transfer examples.

Critical strengths

S1. Effective detection with a self-contained inference path.
Sections 4.3–4.4 and Appendix A.3 explain how learned prototypes and an anchor replace external references at inference. Tables 1, 2, and 9 report the highest mean AP across all tested point budgets on three datasets. This combination of reported accuracy and reference-free deployment is a substantive practical contribution.

S2. Evidence of gains beyond dense query supervision alone.
Table 4 includes a query-only baseline trained with dense masks, with full TCPL improving mean AP by 6.21 and 10.73 percentage points on Aircraft-14 and Vehicle-74, respectively. Table 5 additionally reports a 4.94-point Vehicle-74 AP improvement from adding KD. These comparisons support the utility of the training framework beyond simply having dense labels, while leaving the attribution to individual mechanisms open.

Critical weaknesses

W1. The evaluation supports different degrees of scene generalization across datasets.
Appendix B.1 describes Vehicle-74 as non-overlapping patches from one AeroRIT flight-line image. This evaluates transfer to held-out patches within an acquisition, rather than independently establishing robustness across acquisition conditions. Aircraft-14 uses acquisition-scene grouping, and Plane-400 uses original-image grouping, so this concern does not invalidate all transfer evidence. Nevertheless, it limits the interpretation of the substantial Vehicle-74 gains as evidence for the paper’s broader open-scene motivation. Providing the spatial split map and qualifying the dataset-specific conclusions would address the interpretation; a cross-acquisition claim for Vehicle-74 would require corresponding evidence.

W2. The central multi-support consensus mechanism is not independently isolated.
In Tables 3 and 10 and Appendix B.4, removing TCSV simultaneously replaces attention retrieval with mean pooling and replaces the multi-support teacher with a single teacher. Consequently, the observed difference cannot distinguish the contribution of consensus from that of retrieval. Figure 5 provides an illustrative example, while Table 5 tests KD against no KD rather than multi-support against single-support distillation. A controlled comparison retaining attention retrieval, support exposure, and optimization budget while changing only teacher aggregation would resolve this attribution issue. This matters because consensus-guided transfer is a central contribution, although it does not negate the complete model’s reported performance.

Questions for authors
  1. Regarding W1: Could you provide Vehicle-74 patch coordinates and the train/validation/test spatial layout, including any separation between subsets? This would clarify whether the results should be interpreted as spatial generalization within one acquisition and whether nearby patches could inflate the apparent transfer performance.

  2. Regarding W2: With attention retrieval and the remaining training setup fixed, how does a multi-support mean teacher compare with a single-support teacher under matched support exposure and optimization budget? A consistent advantage would materially strengthen the evidence for the paper’s central consensus contribution.

4 comments
AuthorsOpenReview ↗

Vehicle-74 Spatial Split and Controlled Evaluation of Multi-Support Consensus

Thank you for the detailed feedback.

To Q1:

Vehicle-74 comes from a single AeroRIT flight-line image. We partitioned the image before cropping. In original-image pixels, the training, validation, and test regions span [7,1735)[7,1735), [1735,2247)[1735,2247), and [2247,3975)[2247,3975) along the horizontal axis, with 32, 10, and 32 retained patches, respectively. Within each region, we cropped non-overlapping patches starting at y=53y=53 with a stride of 256 pixels and padded edge patches to 256 × 256 after cropping.

The 512-pixel-wide validation region separates the training and test patches, so they are never directly adjacent. We also checked for vehicles crossing subset boundaries and excluded the affected patches. No source pixels or vehicle instances are shared across subsets.

To Q2:

With attention retrieval fixed, the multi-support mean teacher increased mean AP by 4.71 percentage points on Aircraft-14 and 5.71 on Vehicle-74 relative to the single-support teacher. Mean AUC increased by 0.24 and 0.59 percentage points, respectively.

At each training step, both settings used the same four supports for each query, and all four predictions contributed to the detection loss. The single-support teacher used the prediction from the support previously selected least often as teacher for that query, with ties broken by support ID. The multi-support teacher averaged the four predicted probability maps. Support exposure, supervised losses, and the optimization budget were identical.

Table 1. Retrieval and teacher aggregation (%, mean ± SD)

Retrieval + teacherAircraft-14 APAircraft-14 AUCVehicle-74 APVehicle-74 AUC
Mean pooling + single42.01±1.6197.76±0.3155.40±3.2792.24±2.22
Mean pooling + mean38.53±5.0997.47±0.2655.96±1.9892.19±0.78
Attention + single38.25±4.5897.45±0.1056.81±2.7093.14±1.76
Attention + mean42.96±4.8097.69±0.4162.52±1.5093.73±0.95

To W1:

We added the Vehicle-74 spatial split map and per-patch coordinates and clarified that its results measure generalization to a held-out spatial region within the same acquisition.

To W2:

We added the attention + single-teacher and mean-pooling + multi-support-teacher settings to Table 1, allowing the contribution of teacher aggregation to be evaluated with retrieval held fixed.

Dear Reviewer, with less than three days remaining in the discussion period, we would appreciate your review of our rebuttal and your feedback on whether it addresses your concerns. Thanks again for your time and help in improving the quality of this paper.

I apologize for my late reply, as I have been very busy recently. Thank you for your detailed response. The authors have addressed some of my concerns, and I believe this is a solid piece of work. I will maintain my score of Weak Accept.

Thank you very much for your time and kind assessment of our work. We would be deeply grateful if you might consider these improvements and clarifications in your final evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We have revised the manuscript and added experiments addressing the main concerns:

1. Supervised comparisons: Added DeepLabV3+ with matched dense masks, spectral inputs, splits, and model selection, including evaluation on both San Diego scenes.

2. Generalization evidence: During review, we annotated 167 additional test patches from 13 images at 10 unseen airports for frozen-model evaluation. We also provided Vehicle-74’s spatial map and coordinates and clarified each dataset’s evaluation scope.

3. Consensus contribution: With retrieval, support exposure, supervision, and training budget fixed, multi-support teachers improved mean AP by 4.71 and 5.71 percentage points on Aircraft-14 and Vehicle-74. We added identical-query support-sensitivity and per-source-image analyses.

4. Method clarification: Revised the Introduction, provided actual text examples and a category-only ablation, and clarified automatic description generation and FSS supervision protocols.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.